The GCC Pod Ep 3: Making AI Work, Organizational Failure, Non-determinism, and Context

So over the last two years, I think AI models have crossed into this very strange sort of threshold. Right. I think on benchmarks they reason and they can write and see and they can code and they can even plan, I think better than most humans even. But when you walk into very large companies, you kind of realize that very little has actually changed. So the conversation I think isn't about whether AI works. The conversation isn't about that anymore. It's very clear that it does work. I think the real question is why hasn't that intelligence translated into genuinely real enterprise outcomes? And the models are so strong, but the results are so weak because we keep hearing that models are getting exponentially better. But if that's true, why are so few AI systems actually running? Why?

There's an MIT report.

Exactly. That says 95% of AI systems. So talk a little bit about that.

So I think the first thing is it's true that models have improved a lot. But one thing also ask myself is model improvement is very different than model deployments in them in the enterprise. So enterprise have a lot of context. How to bring that context into the models. That's not an easy thing. I mean, of course we, I mean so think about like, so I was thinking about our own organization and we use this third party system to do annual evaluations, right? And someone who built that system which we are using when they were built, of course that was a SaaS era, right? They may have thought of like they would build it once and then they would sell it to thousands of organizations who are going to make a lot of money. And that's true, that's exactly what SaaS was. But then you know now that the way the templates that platform is giving you for the valuation is right. And you know that's wrong. That's not gonna, you don't want to plan your next year because what's, what's, what's that temp. But based on what the template ask or should ask you to provide and your people to provide onto that on that platform. So that's, the systems are very deterministic. They were designed in a very, in a very deterministic kind of world where they were like you knew what's the input, what's the output. And now you can, I mean I can imagine I can do 100x better if I have more AI based systems. Right, but how can we adapt to that? One thing I've always is like if you are expecting your existing teams to just run the same process into the new platform that's not going to happen. They have a different job to do, for example, in terms of evaluation. So HR has a lot of on their plate and they can't do this now. They can't now work with you to help you modernize that HR stack, that evaluation stack or whatever. So I've been part of some enterprises and I like that idea of they build this change management team, very small team dominated by the operations people. The idea that the great ideas would come from the technologists. I haven't seen that great idea from the technologists. Bring operations people into that change management team and then allow the technologists to work with them to come up with a platform based on the vision that your leaders have for the business and then build very personalized system. That is a way to go rather than getting your head just stuck into the era of the SaaS where you think that you would ask someone and they would give you that platform and then they would do some tweaks and it would work for you. So thinking about AI as a SaaS is wrong. So that's my opinion. So I've seen organizations and I've been working with some of them that if you deploy hyper personalized systems and primarily chose to work with the operations rather than with the technologists, the outcomes are there, models are already there. Of course execution is always harder. So I think that's what I've learned. I mean one answer of this thing where I'll say is bring operations people, bring people who are in the trenches closer to the AI rather than bringing technologists building the system for the operations people. And now the cost of building software has gone very down. So what we really need to find out is what suits the enterprise workflows to your own business workflows.

Right.

Just under that. So we were working with this very large hospital and they were trying to implement this system and there wasn't any adaption. So at the end of the day there were lots of forms and no one was using those. I mean at the end of the day so it was just like. So I mean, and the way they were using it, like for example as an active system of record but there wasn't any intelligence into it and it was very hard for the people to adapt. All we did was and it's easier. All we did was we just asked them how you organize your workflow and build a conversational UI on top of it voice and all that. The adoption is 100. I mean let me tell you, it's far harder to come to the office than coming to the office and not using a system that's especially being procured for you. Like, people want to use technology. It's not like somehow the neural or biological neural weights are not wired to use 2000 or 200 screens.

Yeah, yeah.

That someone else has designed for you with a notion that they would build it and then would sell it to thousands like you.

And I think at some level, you know, I agree with that. But I think also at some level, like the failure in enterprises or large companies for that matter, I think it stops being technical at some point and then it starts becoming structural. Right.

So it's sort of like, I think technical breakthroughs. It was never technical, I don't think. I mean, what? Technical? No, it was always structural. It was always somehow. I mean, your sales tell you and the vendors tell you that's what you need because your competitor needed this and you just bought into that idea and then you deploy that system. You never go back and ask your operations, look, what is the single most important thing you need to have that makes your life 100x easier in doing evaluations, let's say, or doing sales and marketing and whatever. Right. So you never ask that. So someone else tells you, look, and that is good. I mean, it's not bad. I mean, consultants, you hire consultants for that. But at the end of the day, I think people who are in the trends know a lot more than someone else telling them how to run their business. It's always structural, so. Yeah, it's always structure. It's not technology.

Yeah, yeah. And I think, you know, so if we, if we kind of stem out of that, like, there's. There's a lot of optimism. But then in a lot of cases, there also. There's also a lot of confusion. Right. So, for example, you know that there are conversations that are happening today about, you know, things like agents running companies. Right. So why, if we could go a little bit deeper into this particular part of the conversation, why does that promise break down? Like, I want to pick your brain on that a little bit more.

I mean, it's really. I mean, what does it mean? Agent. Companies run by agents. So we use a lot of agents. I use lots of agents. For example, I have a cms and I have an agent on top of it. Then I use different engines for project management that we built on top of our tools. So I use a lot of agents. So, for example, think about that. Every day I'm leading a product, and every day I just extract a report from the agent that traditionally had to assemble from just looking at the dashboard and lots of inference based on that version. And that made me harder to do it every day. So I used to do it, let's say twice, once in two weeks, but now I do it daily. So. But the way the adaption comes in when it's so easy to extract that go through that assistant, which works with the voice, which works with natural language. I use lots of voice AI and then so agents are helping me a lot there also in sales, my sales agents, they are helping me a lot. But the challenge, what I find is, I mean, when you ask. So the way I use agents is like I always. So once the agents are there, once you have developed, the agents always work with the ChatGPT or Cloud, whatever LLM you're using for general purpose. So I asked them that give me a prompt for my, this agent and I try to give it a context. I give that context and then it extracts me the board. So I use the. I create a prompt from a different language model and then I use it on an agent. Now the better job I do in giving the context and preparing a prompt, the better the results are. Sometimes it gives me the reports of holy shit, I haven't thought about it. And sometimes it just, it's like how to make it think I don't need this. I mean, this is bullshit crap. And when you really look into it, it's all about context. And if you just extrapolate on top of it, it also means agents would work the way they are intended to. It's the effort we need to put into give the organizational context in sales and marketing. In my mind, for example, it's not one thing that goes on in mind. So for example, if I'm doing engineering in engineering and now I'm doing, for example, I'm talking to someone about sales. So that engineering context is in my head, right? Generally that context in deterministic software systems, that context, you didn't have the access to that sort of context. All you had to do was you had to ask to your engineer what's going on and you could. So really it's all I currently I have this view, like the better context you have, the better your agents will perform. And we should do some deliberate effort to give context to the agents. I've seen some of the people, they are very good linguistic, they have very good command on linguistics. They're not engineer. And somehow when they talk to ChatGPT, they somehow get results or prompts in such fashion that. And when they use those Prompts into agents. Agents, I mean human can't extract that kind of report that agents does. But really you need to command how to ask the right question, how to tell the agent, how to give the context of the situation.

And so I think two things that come to mind here, like one you already spoke about, which is obviously context, right? Because with talking to any agent or LLM, it's garbage in, garbage out, right? So as good as your prompt will be, as good as the context will be, that's as good as your answer will be, right? But I think it also naturally like brings me to data, right? Everyone says, and this is established, that data is the fuel for AI. And this is something that you and.

I and we've been, we've, we've grown up listing this phase. Data is the new aisle. New aisle. And yeah, yeah, there you go.

Right. So, so that, that kind of brings me to asking, you know, why isn't synthetic data or for example automated data, why isn't that enough for complex reasoning in the enterprise?

So I mean, I'll come back to that because that's it. So in data there are two parts. One is the data you already have in your organization. Does that make sense? First of all, I mean, how much data do you have around sales and marketing and engineering, whatever you do, where is that data and does that make sense? The real effort goes into how to make that data AI ready. The real data for very large organization data exists in house, right? So we need to make sense of that data. That's number one. Now in terms of data creation, let's suppose if, in terms of, if think about again, you want to develop a voice AI for one of your platform, right? First you need to think why you need to do that. Are there any existing agnostic platforms that you can deploy? You should always go for that. But if you don't, then you thought about like, oh, you want to build your own YCI stack, okay. Then the thing is like just even covering accents in voice AI is a very tall order to live up to. So Southern American accents, Northern British accents, Hindi, Urdu, I mean there's so many, just in so many accents there. Now if you want to make sure that your voice AI is good in particular accent or particular language, you need to create a lot more data, right? That probably you don't have in GPT in, on the Internet or even in house. So that's, that's also important. It's very, that's very basic. But also take another use case if you are doing, let's say if you are researching on quantum physics, right, and you want someone who can create a lot of data around quantum physics or astrology and what it requires is like let's say hiring an astrophysicist or quantum biologist from Oxford. That's the second kind of data that you need to train the model. So yes, data is a borderline. But I always, I think enterprise, I mean I was talking to this very large organization and they were thinking about data creation synthetically and key questions I figured out like I mean we don't need any synthetic data creation. All we need is just to make sense of the existing data. The organization has been for more than 100 years. They have so much data. There was so much data even in Excel. We just needed to make sure how to make sense of that data and push it through to the language model and make sense of it. So and by the way, data is very tightly associated with the context. Data is context.

Yeah, exactly, exactly. And you know, that makes me then come back to the thing that we mentioned in the start where you know, you brought up the MIT report and so like why really like why do long term AI strategies seem to like not live up to their potential? Like they fail so quickly.

Long term strategies just remove the word AI into it. I mean I was discussing with Adil and we were Planning of course 2026 and we were planning what to do and we were. I mean I'm. When I started building coordina I had this huge I read book of John Doerr okrs. Google was a thing back then. It wasn't, it's still an empire. But I mean we used to follow a lot. I mean there were books, how Google works and okrs Very important thing. And we used to. So I had this obsession with okrs. So I used to make lots of OKR for the next year and all that. So one thing very important other just said to me a couple of days ago, like think about how can you plan something for a quarter materially, right? Because that's the maximum window that you have to do something. Things are moving so fast. And that's what I also heard from Aravind Perplexity. But it's very hard to think beyond first quarter or one quarter, right? You can't think beyond. I mean you have that macro level overview for example. If you talk about perplexity. Aravind had this macro view that he wants to answer everything. I mean he wants to maximize truth. It's having an answer of whatever. The question you ask is a very Tall order, live up to. That's a macro level thing. But that's not you are going to do, you are going to plan your quarter on. That's a vision, right? But on the quarter front when you talk about AI, AI is not a vision for your company that you say, okay, you have this locked up vision of making digital doctors, right. If you are a hospital, let's say that's not a vision, that's not. AI should not be associated with the vision. AI is a tool, then you want to increase performance or solve certain problems. So think about healthcare. If you think about healthcare, I think in US the per capita cost, the per capita spend on healthcare is somewhere around $14,000 and 40% of that cost is admin cost right now. Which means there's a lot of admin cost that can be replaced with AI. Okay. When you are planning your quarter you can say okay, what kind of tasks you can replace that are very admin oriented. Right. And that are admin in nature. And then you can play with AI in first quarter. That is a way to plan with AI rather. I've seen a lot of people, they say that, I mean I was going to transform biology and what I see is like their admin work is still very manual. So that's the way to look at it.

And I think you know the goal there then making it less about like you said, like I've also been in conversations where AI is part of people's vision. Right. And I think that the focus needs.

Exactly, yeah, shouldn't be, that shouldn't be the part of vision.

Yeah. And, and, and so the focus needs to shift towards, you know, the having the capacity to adapt because you know you, you mentioned quarters and you mentioned essentially you're looking at a horizon, right. You're looking at a window. And so you should worry more about having like fast learning loops. Right. And, and the capacity to adapt again. So I think what you're saying is exactly on point and if this is true, which again in my opinion it is, what does enterprise software actually look like for example 2026, like in this new world? Like what does it look like?

You know, when I started my career, so I was an engineer that I was part of pre sales and all that one thing I learned was that I mean when we used to talk to the customers they used to say like I mean it's harder to use. I mean the workflow, our workflow is different then even I work with some of the consulting businesses like SAP and Oracle and all those and Their consultants like, like, like as that supreme leaders, they used to come look, that's how you run the business. And the guy who's been running the business from last 25 years and telling them like that's the way to run the business, right? Really what it was like look, we have built this because we could scale it for businesses like yours. We have built this, you have to adapt to it, right? Even I mean and I found it very hard and businesses implementation consultants and whatnot, those kind of businesses, whole, whole job was just to customize the core software as per customer needs. That's what for example think about SAP consulting businesses or Oracle Consulting. That's what they do, right. I think I, I think enterprise software shouldn't look like that anymore. Enterprise software should look like how you run your business, right? That's your, that's your advantage. If you are deploying the same CRM that everyone else has deployed, one thing you have for sure is like think about sales CRM. There are contacts, there are ds, there are these follow ups, there are these agents, there are ABM agents. Someone has designed it for you and then of course have given you this idea of like some local platform built in within the CRM so you can customize some of your own workflows. But eventually someone is telling you how to do sales, right? And then you start focusing on other things because someone has done it for you. I think that is wrong. I think that Debbie Dutch recently seen one of his research. Like humans are unimaginably different. And so if humans are so our businesses are. That's where the ideas come in, right? The more you automate your systems, the lesser ideas you would have. Ideas come. I mean the best of the ideas that I've seen that I had come in, it was never for one particular thing. The reason I added so many ideas and I implemented some of them in coding was like I was doing marketing, sales, engineering, coding, all that and then I was just able to cross match. So as your systems, systems should be one consolidated system that should be based on AI, that should be the think of like pool on top of which you need, on top of it you need to build use cases. So think about like you have this AI pool which is very non deterministic, right. But that's one, think of it like one chat GPT for your organization. I'm just giving a reference. You should have your own, let's say model if you want. So for that one model that should be your model should be your active system of record and on Top of which you can. On top of it. You need to develop workflows and systems. Why? Because the learning is then across the department. So that's. And the system should be hyper personalized. Don't let any sales guy tell you how to do your sales and marketing and evaluations and all that stuff and procurement and whatnot. Right? What really I think going forward, any business need is coming up with truly original ideas. And that would only come when you have an active loop. When you have active loop going on in the organization that touches every department. And the idea would generate from that. From that loop.

And when you do talk about generating ideas and having this learning loop, I think it's also important to touch on in the remote first world that we are in right now. This almost sounds counterintuitive in a sense, but with all of this technology, keeping what you just said in mind, why does physical presence then still matter? And if it does, then how does that fit into this entire picture of software and learning loops and ideation and all of those things?

Adoption is hard. Physical presence, Yes. I mean, think about it. If I have to do something, let's say, if I have to ask something about, let's say our platform hyper currently, what I can do is I can walk in straight into the hall and talk to Malik, let's say, and ask why you've done what you've done, right? Because the idea that came into my mind was very instantaneous. I haven't thought about that idea. I don't have any control in my mind when to. I mean, I can't stop ideas coming into my mind. So whenever it comes, I just want to shoot it immediately. If I'm not, for example, if I'm around now, the other way is like, okay, you just write that idea and you can discuss it the other day. So if you give me a system that for example, can just let's say whatever ideas that I've. Whatever things that I want to interact with. Why do you connect with people? Because it's a sense of connection, but also sense of where you want to brainstorm out things, want to get things done. So if you give me the idea that have a persistent memory that does that execute those ideas on my behalf. I did this podcast with Rafia where this thing is called System 3, a meta layer outside AI systems that would just do things on your behalf. So I'm doing this work on my home, did all that stuff and put it into that model and I moved on. But that model would do is it would be just Communicating with those people, talking to them or their agents for example. And that conversation would happen and then value would drive. Currently our models are not there. The tool used for model is hard. I mean I see models, they're finding it very hard to do even shift from one tool to another. There's a company called, there's a model, Korean model called hyperthlover. They've done a very good job in tool use. But that's, that's still in the evolution. Once we are there, where agents then for example can do a lot of work on your behalf. Then yes, you can tell to your boss, look, I'm off, but my agent is coming in the office. He's in the office or she's in the office. Yeah, I mean they would work, but I'm off. Yes, that would happen in next three, five years. But currently we're not there. So. Yeah. And it feels alive. Connection feels alive. I mean, you know, I mean I'm a big fan of on prem work. I hate remote. Most of my teams are on prem, even though probably in some of the use cases, some of the areas where they can work remotely, but I don't allow them. So yeah, I think models are not there. Once there from performance standpoint, there won't be an issue maybe in next three to five years. So yeah, but currently, yes, models are not there. And when I say models, probably in this benchmarks, they are there. But in terms of tool use, in terms of fully context switching, they're not there as human minds are.

Yeah. And just to add to that, I also think that, you know, proximity now has become less about control, which it used to be obviously and more about this context, being able to carry that context and being able to have a tight feedback loop and all of those things. But so you know, if you had to, let's say if you had to like summarize all of this, what, what would you say is. Because we can keep talking on this for like hours and hours, but what would you say is like the core mismatch that's sort of causing AI to underperform in enterprise? Like what's, what's that core thing.

In enterprise or any other business? Like I want to be very enterprise. I would ask myself like why, what are the things that I've achieved with AI in the organization and what I couldn't. Right. So in what I think is AI is there. I think even if there is no further evolution in AI, you can transform your workplaces and factory flows. But what you've already got that's what stops, right? So that's already there. The thing I work, to me, what I find is people. I don't have to say that, but yeah, that's true. People who were younger, 23, 24, 25, just out of college, they are finding it it's far more easier to roll out AI systems with them in the organization than the people who have somehow wired into pre existing ways of pre existing time, like when there wasn't any around. So as in software, we are software business. So building software, I mean they would come lots of arguments why it's not a good idea to build software using let's say hyper cursor. They would say that's. I mean, no, but the truth is people are, I think it's so, so, so hard to rewire to unlearn your neural weights. So if you are in your 30s, it's so hard to earn neural weights. I talked to a lot of CTOs and they are finding it so hard just with the adoption of cursor. Right. I don't know why, but one thing I'm sure it's a people problem, it's not a technology problem, right? So what I've done is I brought in people in some of the key areas where people are more receptive to new ideas. Just want to know how to get things done rather than those who are just wired up. If someone tells me, look, I mean this, I mean, I can tell you that I have some of the engineers, a good friend of mine in my company in coding and they say like, look like, I mean they have this internal competition going on with AI saying like, I mean, I'll prove you. I mean this, I mean, as if AI is some human, they cannot perform better than me. That's not an area or a battle you want to pick, right? You don't want that. I mean it's a tool. So I think it's always the people problem. And also if you are a leader, your own adaption, how your own adaption is. So you know, I mean how many AI tools we use, right? Just think about these are meta glasses. Just think about these are meta glasses. And now it's what I mean it has improved. They don't expect this one bulk improvement in your overall things. They're very small improvement that you can do with AI that would just fundamentally transform the way you work. So it has a native integration with meta AI though meta AI sucks, but still. So I'm driving a car, I'm running, I'm doing whatever, right? So all my Ideas. Whenever I get an idea, think about I'm listening a podcast, right? I've got an idea, right? And even say, hey meta, can you set up this idea into this thing? And I can immediately register my idea into meta and then you can pull it off the next day. And in that pulling off you can integrate that into your notion where that idea is available, right? And even directly with the meta I that idea as well. This has improved my productivity a lot. Second, for example, whisper flow, I use whisper flow. I can't think of writing email with these meat sticks. I can't think, I can't go back because when I write email using whisper flow, it's just better, right? It's just better because the articulation is better. And one thing people say that you can use ChatGPT, they have the voice and you can just copy the text to the email. But really this copying, you can do that? I mean, but, but then when I, when I go straight to the mailbox and write using whisper flow, somehow my articulation is very concise and better. I don't know why, but again, yes. So you have to upgrade your toolkit as a person also for all the retail, using all the retail product. Once that's there, then also think about if that's there then you can then you have the ability to generate ideas within the enterprise too. And then of course hiring the right people, bringing the right people, relatively younger people, less than age 30. And it's not like, I mean there are a lot of people about 30. But I'm saying is like we have the, we have a lot to learn. So let them learn on a new stack and they can adapt and roll out. I mean great. If they don't know the any, they don't know how to, how the work was done before AI, that's great because that's their starting point.